Adaptive Variable Resolution 2.5D Lidar Mapping for Dynamic Environment Perception
DRDO · Transportation & Logistics · Software
Public data, numeric targets, a measurable memory claim and a genuinely novel representation make this among the cleanest statements on the portal — spend your time on the variable resolution grid rather than the segmentation network, because that is where both the difficulty and the credit are.
What it actually is
A LiDAR throws out millions of points a second, which is too much to process live, but flattening it into a plain 2D map throws away the height information you need to see a kerb or a low branch. The ask is a middle path modelled on how eyes work — full detail close to the vehicle where safety depends on it, and progressively coarser further out. The map should also know what it is looking at, not just how tall it is.
What to build
A perception pipeline turning raw point clouds into a semantically labelled elevation map at non-uniform resolution: a segmentation network classifying every point into drivable terrain, static obstacles such as walls and poles, and dynamic objects such as pedestrians and vehicles; a variable resolution grid engine projecting those classified points into a 2.5D structure at the resolutions the statement specifies — roughly five centimetre cells within ten metres coarsening to fifty centimetre cells out to a hundred — handling the alignment and data loss problems that arise when cells of different sizes meet; a real-time visualisation colour-coding terrain and object classes; and a benchmark reporting frame rate, classification accuracy broken down by distance band, and memory footprint against a uniform high-resolution baseline.
Smallest thing that wins the room
Play a driving sequence live with the foveated grid visibly coarsening outward, a pedestrian tracked in the fine zone, and a memory counter showing your map using a fraction of what the uniform 3D representation costs at the same frame rate.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 16% of the 226 · #191 of 226 by expected field · reaches the 500 cap
Busier than most. Expect several teams to arrive at the same obvious solution.
Why: defence, intelligence and space bodies drew small fields.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
4/5One of the cleanest statements in this range — public data removes all sourcing risk, the targets are numeric so success is measurable, the memory reduction claim is objectively verifiable in the room, and the foveated representation is a real idea rather than a repackaging of a solved benchmark.
Feasibility
4/5Large labelled LiDAR datasets with per-point semantic annotations are public and standard, the segmentation architectures the statement names have mature open implementations, and no sensor or vehicle is required — the entire project runs on recorded data on a single GPU.
Innovation scope
4/5The foveated variable-resolution grid is a genuinely non-standard representation and the statement explicitly flags the hard part — a data structure that handles varying cell size without alignment errors or data loss during projection — as unsolved, so the core contribution is yours to invent.
Clarity
5/5Remarkably concrete: three primary tasks, actual resolution numbers and radii, named candidate architectures, the specific failure mode to avoid in the projection, the visualisation requirement and the three performance metrics are all stated in about two thousand characters.
Effort
HeavyTraining a point cloud segmentation model, designing and implementing a variable resolution grid structure, building a real-time visualiser and running a comparative benchmark is four substantial pieces, and the grid engine will take longer than the network because there is no library that does it.
Demo-ability
EasyA live map rendering with visibly varying cell size, colour-coded semantics and a memory counter ticking against the baseline explains itself in seconds and shows the actual claim rather than describing it.
In its favour
- Green flag: The memory reduction against a uniform grid is objectively measurable and demonstrable live, which gives you a hard number to win on rather than a subjective quality claim
- Green flag: Public per-point-labelled LiDAR datasets remove the data risk entirely, and they come with established evaluation protocols so your accuracy figures are comparable to published work
- Green flag: The statement identifies the genuinely hard part for you — alignment and data loss across varying cell sizes — which means you can go straight at the interesting problem instead of discovering it in week two
- Green flag: The resolution figures are given, so your grid design is anchored to a stated requirement and cannot be criticised as arbitrarily chosen
Against it
- Red flag: Segmentation is the easy half — mature open models will get you most of the way — so a team that spends its time on the network and bolts on a naive grid has skipped the actual contribution
- Red flag: Alignment across resolution boundaries is genuinely fiddly: a cell straddling the fine and coarse zones will either double-count or drop points, and the artefacts appear exactly at ten metres where safety matters most
- Red flag: Semantic labels in the public datasets come from a driving context with class taxonomies that do not necessarily match a defence sponsor's terrain categories, so map them explicitly rather than adopting them silently
- Red flag: Real-time claims need to be measured end to end including the projection, not just the network's inference time, and the grid update is often the slower half
What you will be writing
- MinkowskiEngine or SPVNAS sparse convolutional segmentation
- PointNet++ point cloud feature extraction
- quadtree-backed variable resolution elevation grid
- SemanticKITTI / nuScenes labelled point clouds
- Open3D or rerun.io real-time visualisation
- memory footprint benchmarking against uniform voxel grid
- LiDAR perception
- Autonomous navigation
- Spatial data structures
Prior art to read before you start
point cloud semantic segmentation · variable resolution occupancy mapping · real-time perception memory optimisation
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.